See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/261262060

Genetic Algorithms: A tool for image segmentation

Conference Paper · May 2012

DOI: 10.1109/ICMCS.2012.6320144

CITATIONS READS

22 4,180

3 authors, including:

Alaa Sheta Malik Sh. Braik

Southern Connecticut State University 61 PUBLICATIONS   1,607 CITATIONS   
196 PUBLICATIONS   2,615 CITATIONS   
SEE PROFILE
SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Automatic Detection and Classification of the Ghor-area Plant Diseases View project

Model-based industrial systems View project

All content following this page was uploaded by Alaa Sheta on 21 June 2014.

The user has requested enhancement of the downloaded file.

Genetic Algorithms: A Tool for Image Segmentation
Alaa Sheta Malik S. Braik
Computer Science Department Electronic, Electrical and Computer Eng. Dept.
WISE University University of Birmingham
Amman 11947, Jordan Birmingham, Edgbaston, UK
Email: asheta66@gmail.com Email: MSB158@bham.ac.uk
Abstract— Genetic Algorithms (GAs) are increasingly being ex-
plored in many areas of image analysis to solve complex opti-
mization problems. They rapidly gained acceptance in the scientific
community as powerful statistical search method whicj allows us
to consider the segmentation problem as an optimization problem.
In this paper, we propose the use of GAs in an integrated manner
with traditional image segmentation techniques to provide an efficient
segmentation and edges detection for selected natural images. The
developed experimental results are compared with the results of other
known existing segmentation algorithms such as K-mean clustering,
and global threshold methods. The proposed method is capable of
achieving a satisfactory results. Accordingly, the GAs based image
segmentation method will definitely help in solving various complex
image processing tasks.
Keywords— thresholding, image segmentation, genetic algorithms,
edge detection, optimization
I. I NTRODUCTION
Image segmentation is an essential step for many advanced
imaging applications, and considered one of the hard problems
of image recognition and understanding. In case of analyzing
and understanding the information in an acquired image, it is
necessary to extract the area in which the objective material
is included. Moreover, depending on the problem domain or
application, there are several types of images that could be
processed and analyzed such as, light intensity (i.e. gray scale),
color, range, Infra-Red (IR) [1] and SAR images [2].
Many range image segmentation algorithms was proposed in
[3]. These segmentation algorithms were tuned using number
of heuristic algorithms such as GAs to provide exact results on
a given group of images [4]. Segmentation methods frequently
use thresholds to decide whether two pixels are in the same
region or not. However, the thresholds negatively affect the
image segmentation results often in accuracy. Therefore, the
tuning values of the thresholding algorithms must be set up
carefully [3], [5]. A multilevel thresholding method which al-
lows the determination of the appropriate number of thresholds
as well as the adequate threshold values was proposed in [5].
This method combines both genetic algorithm and wavelet
transform. In [6], authors presented a new distributed image
segmentation algorithm structured as a multiagent system
composed of a set of segmentation agents and a coordinator
agent.
A. F. Sheta is a Professor of Computer Science and Engineering, Computer
Science Department, The World Islamic Science and Education (WISE)
University, Amman, Jordan. He is on leave from the Electronics Research
Institute (ERI), Cairo, Egypt
Sultan Aljahdali
Department of Computer Science
Taif University
Taif, Saudi Arabia
Email: aljahdali@tu.edu.sa
Despite the large number of segmentation techniques
presently available, no general methods admit a unique so-
lution because contest may take part in the decision phase.
It follows that general solutions are not possible, and each
proposed technique can be used only to solve class of problems
[7]. Consequently, computing such an image partition is a very
high complexity problem, hence, development of a unified
method to image segmentation is still an open research area
which significantly in need for more effort [8].
This research proposed a new method for image segmenta-
tion problem using Genetic Algorithms (GAs). Image segmen-
tation problem is formulated as an optimization problem based
GAs. The contribution of this work is the use of GA to build
a reliable and accurate method to tune a set of parameters for
a traditional filter. One reason for using this kind of method
is mainly related with the GA ability to deal with complex
search spaces in situations where only minimum knowledge
is available. Moreover, the proposed GA method for image
segmentation is compared with other traditional methods in
the literature. The applicability of the proposed methods has
been tested on a large class of images; the obtained results are
encouraging, performing well in all of the proposed problems.
II. W HAT IS GA S ?
Genetic algorithms (GAs) are adaptive search procedures
which were first introduced by Holland [9], and extensively
studied by Goldberg [10], De Jong [11]–[13], and others. One
of the most successful areas of application has been the use of
GAs to solve a wide variety of difficult numerical optimization
problems. GAs complement existing optimization methodes
nicely in that they require no gradient information and are
much less likely to get trapped in local minima on multi-
modal surfaces. In addition, GAs have been shown to be quite
insensitive to the presence of noise [14].
• In GAs, each possible solution within the population of a
biological individual is coded in so called ”chromosome”
(i.e. individual). A number of chromosomes generate
what is called a ”population”. The structure for each
individual can be represented as a string of characters
which are usually binary digits [9] or real numbers. The
chromosomes share data with other, and each chromo-
some is assigned a fitness score according to how good a
solution to the problem based on a given fitness function.
The solutions are taken according to their fitness values
 and used to construct new solutions by a hope that the • The quality of the segmented results for each parameter
new solutions will be better than the old solutions and a set is then evaluated. If the maximum segmentation
generation is complete. Thus, the whole population moves quality for the current population is above a predefined
like a one group towards an optimal area [9], [10], [15], threshold value or another stopping criterion, then, the cy-
[16]. cle will terminate. Alternatively, if the performance is not
• At each generation, each individual is evaluated and achieved, the genetic recombination operators (crossover
recombined with others on the basis of its fitness. The and mutation) are applied to the high strength individuals
expected number of times an individual is selected for in the population, which yield a new set of offspring that
recombination is proportional to its fitness; the fitness is has better performance.
relative to the rest of the population. New individuals are • The new population is supplied back to the image seg-
created using crossover and mutation operations. mentation process, and thus, the cycle begins again. Each
• Crossover operates by selecting a random location in pass through the loop (Segmentation-Fitness-Crossover,
the genetic string of the parents (crossover point) and and Mutation) is known as generation. The cycle shown
concatenating the initial segment of one parent with the in Figure 1 continues until the maximum fitness or
final segment of the second parent to create a new child. termination criteria are achieved.
A second child is simultaneously generated using the • The fitness function is related to the internal variance of
remaining segments of the two parents [15], [16]. each image segment, and has two weights for grouping
• Mutation provides for occasional disturbances in the the pixels into subsets based on the distance function.
crossover operation by inverting one or more genetic Hence, the pixel intensity and spatial position are used,
elements during reproduction [9], [15]. Therefore, the distance function is based on the pixel
intensity and spatial position to define the similarity
III. I MAGE S EGMENTATION U SING GA S
between pixels as shown in Figure 2.
GA is proposed to explore the solution space, in the sense
111
000 00
11
00 111
that each pixel is grouped into other pixels by distance function 000
111 000 11
00
based upon both local and global already computed segments. 000 f011
111 000
111 00
11
00
11 000
111 00
11
There are two basic issues needed to be addressed in designing 00 11
11 00
00 111
11 000 00
11
a GA for image segmentation, they are:
00
11 00
11 11
00
• 1) selecting an appropriate representation for each poten- 00 11
11 00 11
00
00f1
11 00
11
00
11 000
111
00 11
00 00
11
tial solution;
• 2) formulating an appropriate fitness functions in terms
11 00 111
11 000
000
111 00
11
of the weight. Fig. 2. The graph and the distance function

Almost every image segmentation algorithm contains pa-

rameters that are used to control the segmentation results;
the genetic system can dynamically change the parameters to
achieve the best performance. The image segmentation based
GAs process may be summarized as follows:
• Once the initial population is available, the genetic cycle
begins as shown in Figure 1.
Initial Population
IV. P ROPOSED S EGMENTATION A LGORITHM
The segmentation of a digital image F can be considered as
a natural partitioning of its pixels, based on a distance function
that is defined in F . F of size (M ×N pixels) can be described
as:
F = F (i, j), 1 ≤ i ≤ M, 1 ≤ j ≤ N, 0 ≤ F (i, j) ≤ C (1)

Genetic Reproduction where C is the number of class prototypes that can be selected
New Population Image Segmentation
and Recombination freely, so that, F is a collection of pixels where each pixel
value is proportional to the degree of region similarity. The
Current Population
Segmented Results
pixel values in F are normalized in the range [0, 1]. Pixels with
No values close to one are good candidates for being considered
Terminate? Fitness of Population Segmentation Evaluation
Yes
as edge points. The regions or subsets of F are denoted by:
S = (S1 , . . . , SK ).

Segmented
A. Image Preprocessing
Image
The preprocessing phase of image segmentation includes
image similarity based on the distance function, connected
component labeling, edge tracing, and initial fitness evaluation.
The regions with similar pixels are grouped into regions as
a final result of segmentation or clustering. At the end of
Fig. 1. Flow chart of the GAs based segmentation method clustering procedure, a label is assigned to each pixel that
take a value between 1 and K, such that, each possible region
of F can be represented by a matrix, and it corresponds to a V. E XPERIMENTAL S ETUP FOR GA S
pixel of the subset or partition. F is associated to the weighted For an input image F of size N × M , F (i, j) indicates
graph G as: G ≡< F, δ >. δ is the distance function that links to the gray-level value of the pixel located in (i, j) location.
between f0 and f1 as shown in Figure 2. The proposed procedure of image segmentation based GAs
B. Image Partitioning consists of the following steps:
The segmentation algorithm depends on partition the ver- A. GA Data coding
tices of the weighted graph G = (V, E, δ) into K subsets.
An example of weighted graph of three segments isTshown in The genetic chromosome is coded by a 32 bit binary
Figure 3. As given previously, G = (V, E, δ), Si Sj = Φ, string. The 8 least significant bits identifies the pixel label
of the cluster to which the pixel belongs after the clustering
11
00 procedure. While, the pixel position (i, j) is coded into the 24
00
11
3
most significant bits. Consequently, each chromosome can be
11
00 5 denoted with the order pair as shown in Figure 5.
m=2 00
11 1
11
00
k=11
4
2
2 00
11 b31 , . . . . . . , b8 b7 , . . . , b0

11
00 6
11
00
00
11
Fig. 5. A Chromosome representation

00
11
Fig. 3. Image segmentation and partitioning the graph
B. Fitness Function
Evaluation a segmentation technique is not an easy task,
since the expected segmentation result depends on the ap-
10
01
S0 S1

0
1 1
000
1110
0
1
01 1
0
10
0
10
0 1
0
00
11
1 00
11
00
111
0
1
01
0
00
11 1
1
0
plication. The fitness function of the segmentation quality
10
01 11
1
0 00
11
0 1
000
11
1
0 1
0
varies from image to image. Therefore, the definition of the
1
0
10
0 11
0
00
11
1
0 1
001
0
00
1100
1
11
0 objective function itself can be subject of debate because there
10
1
0
10000
1111
00
11
10
0 0
111
1
0 0
1
0
0
10000
1111
10
0000
11111
01
0
are no single, universally accepted measures of segmentation
1
0 00
111
0
11
0000
1111
00
11 11
0 performance available [7], [17], [18].
1
0 0
11111
000001
0
10
0 The fitness function has been defined on the basis of the
distance function computed between each chromosome and the
Fig. 4. Illustration the bipartition process between two subsets
corresponding region S. The fitness function has been defined
S with two weights characterized by w0 and w1 . In this study, an
and Si Sj = F . Hence, we want to disjoint Si and Sj by
agreement indicator (i.e. fitness function) between human and
removing edges connecting the two parts as shown in Figure
automatic classification has been defined as given in Equation
4). The cut of Si and Sj can be defined as given in Equation
4 [19]:
2:
1 X
K
#agrk
X γ= βk × (4)
cut(Si , Sj ) = δ(i, j) (2) K
k=1
P¯k
i∈Si j∈Sj
where:{P¯k =max(|Segk |, |S|)}, Segk denotes the k th segment
The optimal partition is the one that minimizes the distance retrieved by humans, S denotes the k th segment retrieved by
function (similarity function). A pixel f ∈ F is represented in the machine, |Segk | and |S| denote the corresponding size
a 3D feature space with (if , jf , lf ). (i, j) represents the pixel and #agrk is the largest pixel intersection between Segk and
position, and l represents its gray level. S.
In this research, the grouping of pixels into partitions is
based on the distance function computed for each partition as
defined in Equation 3. C. Crossover and Mutation
|li − lj | The classical single point crossover with crossing probabil-
δ(i, j) = w0 ×
max(li , lj ) ity Pc and bit mutation with probability Pm have been evolved
1 X |lz − lj | in the proposed genetic system. An example of single point
+ w1 × (3) crossover and bit mutation is shown in Figure 6.
|N (i)| max(lz , lj )
z∈N (i)
D. Selection Process
w0 , w1 ≥ 0 and ∀i, j ∈ F, i ∼ j ⇐⇒ δ(i, j) ≤ Φ.
N (i) is a neighborhood of the pixel i. w0 and w1 are the In the GAs poputaion, N × M numbers of chromosomes
weights attached to its feature. Therefore, the ranges of w0 , w1 are generated using random numbers, such that:
are normalized between 0 and 1, consequently, each feature
contributes equally to the distance function. P (0) = {ν(0), ν(1), . . . , νM ×N (0) (5)
 TABLE I
11111111111
00000000000
00 11
11 00 11 11
00 00 11
00
Cut Point

00000000000
11111111111
00 11
11 00 00000000000
11111111111
00
11 00
11 00
11
Offspring1
00
11
Parent1 Crossover T UNING PARAMETERS OF GA S
00000000000
11111111111
00 11
11
1
00 11
1
00 0 0 1 0
00000000000
11111111111
00 11
11 00 11
1
00
1 0 0 1 1

00000000000
11111111111
00
11 00
11 00
11 00
11
Parameter Value

00000000000
11111111111
00
11 00
11 00
11 000000000000
111111111111
Parent2
00
11 00
11 00
11
Offspring2 Population size 50 individuals
00000000000
11111111111
1
00 11
00 11
11
0
00 0 0 1 1
000000000000
111111111111
00 11
11 00 11
1
00
0 0 0 1 0 Gene length
Maximum generation
32 bits
200
00000000000
11111111111
00
11 00
11
00000000000
11111111111
00 11
00 00000000000
11111111111
00 11
11 00 00
11
The Prob. of Mutation Pm 0.01
00 11
11
Mutation

00000000000
11111111111
1
00 11
00 11
11 0
00 1 1 0 1
00000000000
11111111111
00 11
11 00 11
1
00
1 1 0 1 0 The Prob. of Crossover Pc 0.8

Fig. 6. Example of crossover and mutation operations

TABLE II
I MAGE SIZES AND NUMBER OF GENERATIONS FOR GA RUNS
P (0) is the population at time 0. Then, the single point Image size iterations
crossover and bit mutation are applied to the initial population Aish 383 × 801 200
to get another generation, P (1), and so on. The new population Hand 243 × 909 200
House 192 × 765 200
is determined by the selection process such that the new Fish 74 × 390 200
population is better than the old population in terms of its
fitness.
E. Termination Condition The proposed segmentation problem based GAs is treated
The procedure from step a to step d is repeated until the as a partitioning problem. It has been applied to polygons,
convergence condition is satisfied or the number of generations natural, and texture images. The results show good agreement
is reached. The convergence process is based on the total between natural and automatic partitioning. The experimental
variance of the region as shown in Equation 6. results are compared with the results of other known existing
segmentation algorithms such as K-mean clustering [20], and
X
K
V art = σ(k)2 (6) global threshold methods [21]. Segmentation results are shown
k=1
in Figures 8, 9, 10, and 11. It can be seen that the GA
proposed method obtains the most optimal results, but with
where, t is the iteration number, σ 2 is the grey-level variance high computational effort.
of region k, and K is the total number of regions. The loop
continue until |V art+1 −V art | ≤ ǫ, where ǫ is a small defined
value. The GA based image segmentation proposed method is
shown in Figure 7.

for every f ∈ F
if δ(f, lSk )=min{δ(f, gSi ), i = 1, 2, . . . , Kmax }
then,
A single point crossover and bit mutation are applied
to current population;
A new population P (t + 1) is generated.
assign (f, Sk )
update ({µk , σk })
end
end
Increase iteration by one;
repeat until (|V art+1 − V art | ≤ ǫ)
Fig. 7. GA based image segmentation algorithm

VI. E XPERIMENTAL R ESULTS
The proposed segmentation algorithm has been imple-
mented using M ATLAB with Signal Processing Toolbox and
Statistics Toolbox under windows system. Several experiments
are tested to demonstrate the performance of the proposed
segmentation algorithm. The setting parameters of the GAs run
are shown in Table I. A set of four images are used to illustrate
the usability of the proposed algorithm as seen in Table II.
Fig. 8. Results of segmentation: upper left Aish image; upper right Global
threshold; lower left K-Means clustering method; lower right GAbased
method;
The GAs convergence has been achieved in all experiments.
Figure 12 shows the convergence of the segmentation algo-
rithm in the case of hand image. Figure 12 indicates that
the proposed segmentation method based GAs converges to
a solution near the global optimum. The agreement indicators
Fig. 9. Results of segmentation: upper left - Hand image; upper right - Global Fig. 10. Results of segmentation: upper left - House image; upper right
threshold; lower left - K-Means clustering method; lower right - GAbased - Global threshold; lower left - K-Means clustering method; lower right -
method; GAbased method;

of the tested images are shown in Table III. edges running vertically and horizontally. Prewitt operator
needs a threshold value and is very sensitive to noise as seen
TABLE III
in Figure 14. Therefore, the operator is not effective to the
AGREEMENT INDICATORS OF THE TESTED IMAGES
detection of the edges.
Method/γ Aish Hand House Fish
K-Mean 0.69 0.66 0.61 0.63
C. Roberts
Global 0.59 0.53 0.55 0.49 The Roberts cross operator performs a simple spatial gra-
Proposed GA-Method 0.75 0.83 0.80 0.83
dient measurement on the image. It thus highlights regions
of high spatial frequency which often corresponds to edges.
The operator consists of a pair of 2 × 2 convolution kernels.
VII. C OMPARISON WITH OTHER T ECHNIQUES The kernels are designed to respond maximally to edges
For illustrating the efficiency of the proposed edge detection running vertically and horizontally, one kernel for each of the
method, the results are compared to other standard existing two perpendicular orientations. Consequently, only four input
edge detection methods. pixels are needed to be examined to determine the value of
each output pixel, and only subtractions and additions are used
A. Sobel in the calculation.
The Sobel operator approximates the first derivative that Robert operator is very sensitive to noise since it uses such
performs a spatial gradient measurement on the image, and a small kernel, and operator produces very weak responses to
so, emphasizes on regions of high spatial frequency that genuine edges unless they are very sharp as seen in Figures
correspond to edges. Sobel operator consists of a pair of 13, 15, and 16. Also, the operator needs a threshold value
3×3 convolution kernels. The kernels are designed to respond which negatively effects on the result of edges detection. The
maximally to edges running vertically and horizontally, one threshold values of the operators in the case of aish image are
kernel for each of the two perpendicular orientations. Thus, shown in Table IV.
the operator has a problem of identifying edges in three TABLE IV
dimensions as seen in Figures 13 and 15. Furthermore, the T HE THRESHOLD VALUES IN THE CASE OF AISH IMAGE
operator needs a threshold value which negatively effects on
the result of the edges detection (see Table IV). Operator Threshold value
Sobel 0.075000
B. Prewitt Prewitt 0.075000
Robert 0.215700
The Prewitt operator is similar to the Sobel operator. It ap-
proximates the first order derivative, but uses slightly different
kernels. The operator consists of a pair of 3 × 3 convolution The parameter values of the GAs run for the edge detection
kernels. The kernels are designed to respond maximally to problem are shown in Table V.
 Fig. 13. Results of edge detection of the aish image: upper left - Prewitt
operator; upper right - Robert operator; lower left - Sobel operator; lower
right - GA based method;

Fig. 11. Results of segmentation: upper left - Fish image; upper right - Global
threshold; lower left - K-Means clustering method; lower right - GAbased
method;

Convergence of the Algorithm

35

30

25
Variance

20

15

10

Fig. 14. Results of edge detection of the hand image: upper left - Prewitt
5
0 20 40 60 80 100 120 140 160 180 200 operator; upper right - Robert operator; lower left - Sobel operator; lower
Iteration
right - GA based method;

Fig. 12. Convergence of GAs via the number of gerenations: The hand image

VIII. C ONCLUSIONS AND F UTURE W ORKS

In this paper, we used GAs to solve the image segmentation
problem. The problem was treated as an optimization problems
TABLE V based GA. GAs was used to segment an image by using
T UNING PARAMETERS OF GA S FOR THE EDGE DETECTION PROBLEM an optimization function without any threshold values. GAs
based image segmentation can provide more accurate results
Parameter Value
Population size 50
than traditional segmentation methods. The genetic procedure
Gene length 32 provided a faster convergence to the optimal solution. This is
Maximum generation 200 because the sampling strategy allows exploring the solution
The probability of mutation Pm 0.01 space by a strategy that is not biased. The preliminary results
The probability of crossover Pc 0.8
indicate that GA-based methods outperform the traditional
methods in terms of quality. Finally, the obtained results
of image segmentation and edges detection indicate that the

Fig. 15. Results of edge detection of the house image: upper left - Prewitt
operator; upper right - Robert operator; lower left - Sobel operator; lower
right - GA based method;
Fig. 16. Results of edge detection of the fish image: upper left - Prewitt
operator; upper right - Robert operator; lower left - Sobel operator; lower
right - GA based method;
proposed methodes have a high effectiveness on a large
category of images applications. An extension work will focus
on developing hybrid algorithms such as GAs and Artificial
Neural Networks (ANNs) in order to reduce the processing
time and increase the visual quality of the final segmentation
underscoring the advantages of hybrid algorithms.
R EFERENCES
[1] A. Ajlouni and A. Sheta, Image Processing Techniques for Landmine: A
Framework for Detection and Classification. LAP LAMBERT Academic
Publishing, ISBN 978-3-8473-3721-8, 2012.
[2] G. Gao, L. Zhao, J. Zhang, D. Zhou, and J. Huang, “A segmentation al-
gorithm for sar images based on the anisotropic heat diffusion equation,”
Pattern Recogn., vol. 41, pp. 3035–3043, Oct. 2008.
View publication stats
[3] G. Pignalberi, R. Cucchiara, L. Cinque, and S. Levialdi, “Tuning range
image segmentation by genetic algorithm,” EURASIP J. Appl. Signal
Process., vol. 2003, pp. 780–790, Jan. 2003.
[4] S. Chabrier, C. Rosenberger, B. Emile, and H. Laurent, “optimization-
based image segmentation by genetic algorithms,” J. Image Video
Process., vol. 2008, pp. 10:1–10:10, Jan. 2008.
[5] K. Hammouche, M. Diaf, and P. Siarry, “A multilevel automatic
thresholding method based on a genetic algorithm for a fast image
segmentation,” Comput. Vis. Image Underst., vol. 109, pp. 163–175,
Feb. 2008.
[6] K. E. Melkemi, M. Batouche, and S. Foufou, “A multiagent system
approach for image segmentation using genetic algorithms and extremal
optimization heuristics,” Pattern Recogn. Lett., vol. 27, pp. 1230–1238,
Aug. 2006.
[7] A. Brink, “Using spatial information as an aid to maximum-entropy
image threshold selection,” PRL, vol. 17, pp. 29–36, January 1996.
[8] S. M. Bhandarkar and H. Zhang, “Image segmentation using evolu-
tionary computation,” IEEE Transactions on Evolutionary Computation,
vol. 3, no. 1, pp. 1–21, 1999.
[9] J. Holland, Adaptation in Natural and Artificial Systems. Ann Arbor,
MI: University of Michigan Press, 1975.
[10] D. Goldberg, Genetic Algorithms in Search, Optimization, and Machine
Learning. New York: Addison-Wesley, 1989.
[11] K. A. De Jong, Analysis of Behavior of a Class of Genetic Adaptive
Systems. PhD thesis, University of Michigan, Ann Arbor, MI, 1975.
[12] K. A. De Jong, “Adaptive system design: A genetic approach,” IEEE
Transaction Sys. Man. Cybern., vol. 10, no. 3, pp. 556–574, 1980.
[13] K. De Jong, “Learning with genetic algorithms: An overview,” Machine
Learning, no. 3, pp. 121–138, 1988.
[14] A. Sheta and K. D. Jong, “Parameter estimation of nonlinear systems in
noisy environment using genetic algorithms,” in Proceedings of the IEEE
International Symposium on Intelligent Control (ISIC’96), pp. 360–366,
1996.
[15] D. E. Goldberg, The Design of Innovation: Lessons from and for
Competent Genetic Algorithms. Boston, MA: Kluwer Academic, 2002.
[16] K. DeJong, An Analysis of Behavior of a Class of Genetic Adaptive
Systems, Doctoral dissertation. PhD thesis, University of Michigan,
Dissertation Abstracts International, 1975.
[17] D. E. Goldberg, Computer-Aided Gas Pipeline Operation Using Genetic
Algorithms and Rule Learning. PhD thesis, University of Michigan, Ann
Arbor, 1983.
[18] J. Lee, R. Haralick, and L. Shapiro, “Morphologic edge detection,” J-RA,
vol. 3, pp. 142–156, 1987.
[19] V. di Gesu and G. lo Bosco, “Image segmentation based on genetic
algorithms combination,” in CIAP05, pp. 352–359, 2005.
[20] J. Ruisz, D. Honigmann, and H. Pottmann, “Segmentation and modeling
of approximately rotationally symmetric objects in 3d ultrasound,” in
3DIM03, pp. 124–131, 2003.
[21] R. Ohlander, K. Price, and D. R. Reddy, “Picture segmentation using
a recursive region splitting method,” Computer Graphs and Image
Processing, vol. 8, pp. 313–333, 1978.